How can you use real-time time series forecasting where each item cannot be tracked?

Question

Problem . Find the estimated lifetime of the object (for example, the time that will be written later) or the corresponding PDF. This is called the upgrade process .

Restriction : It is not possible to track metadata for each individual object.

Assumptions : forecast errors for objects with a low volume are unacceptable, but the error should decrease with increasing object popularity

Do you have any ideas on how these predictions can be achieved, possibly using sketch data structures (Bloom filters, Count-min sketches, etc.) or sample shapes (e.g., exponentially biased collector sampling ) ? Does a specific random process (for example, the Poisson process) involve an easier task?

A striking example of this problem will be: to evaluate when a user will visit your site or click something, not being able to track the history of each user.